Buildings tend to follow patterns, they are classified by typologies and repeat most of their characteristics. Would it be possible to train a prediction model based on a dataset of technical drawings to generate designs with prompts? Would love to discuss this with the technical folks and hear your thoughts.
I have zero experience in floor planning, but I have read some relevant research articles and I am sure this will be possible if you provide the AI with sufficient examples.
I am unsure how many examples you would need for your dataset, it depends on the complexity of the plans you wish to create.
The bare minimum to begin experimentation would prob be at least 1000 specification : plan pairs.
Rather than technical drawings, perhaps start from building information models (BIM models) that are used to design buildings? These have the building elements (stairs, elevator, primary circulation, secondary circulation, interior/exterior walls, windows) identified. These also let you understand multi-floor considerations (e.g. lobbies, cafeteria, gyms and other office spaces that aren’t repeated floor to floor).
You might also want to add other information about the design goals of each typology to your model. For instance, commercial office space designed to increase interaction in creative teams would include a higher percentage neighborhoods, shared spaces, “town hall” or cafe-style spaces rather than on offices, cubes, and conference rooms.
Another major generator of designs is the architectural program – e.g. what capacity is needed and what are the proximity and communication requirements.
One way to attack the question might be to look at existing efforts in generative design (e.g. Autodesk’s design of their Toronto office Generative Design in Architecture | In the Innovation Zone at AU with David Benjamin - YouTube ), look at their constraints (e.g. workstyle, daylight, communication, etc.) and their scoring of computer generated designs. Then think through what typologies of existing designs would score well and how to abstract them as patterns that endure even when the constraints (e.g. the shape of the lot) change.